Description
- Abstract:
- An essential goal in systems neuroscience is to understand how cell-level properties give rise to emergent network-level activity. An accurate mechanistic understanding of emergent dynamics in neural circuits has the potential to guide novel experimental directions, improve the abilities of artificial intelligence, and accelerate the development of translational neuroscience applications. Computational models of neural circuits are indispensable tools for producing hypotheses on the biological mechanisms underlying emergent neural activity and behavior. However, models which account for the complex biophysics of neurons are highly challenging to use due to: 1) difficult parameter optimization, and 2) high computational cost. This dissertation demonstrates how deep learning can be used to address the computational challenges of biophysical neural modeling and uses these models for previously inaccessible scientific questions. I specifically focus on computational models of the neocortex to understand the cellular properties that impact 1) neural oscillations measured by electroencephalography (EEG), and 2) network dynamics that mimic working memory. For the study of neural oscillations, I use the biophysical modeling framework the Human Neocortical Neurosolver (HNN), a biophysically detailed model of a cortical column that was specifically designed to link cell and circuit-level properties to EEG biomarkers. To address the challenge of difficult parameter optimization, I use the deep learning technique known as simulation-based inference to infer cell-level biophysical parameters that produce specific EEG biomarkers. To address the challenge of high computational cost, I demonstrate how deep neural networks can be trained as surrogate models which approximate HNN simulations with orders of magnitude lower computational cost. For the study of working memory, I developed a novel reservoir computing framework which employs cell and circuit-level properties inspired by the neocortex, and trained it to solve a simplified working memory task. The analysis revealed that NMDA receptors are highly important for delivering task-relevant information to dendrites. Overall, the techniques developed in this thesis establish deep learning as an essential tool for studying emergent dynamics in biophysical neural models, and set the stage for future work on the mechanistic underpinnings of numerous other cortical functions.
- Notes:
- Thesis (Ph. D.)--Brown University, 2025
Citation
Tolley, Nicholas,
"Bridging biophysics and emergent dynamics through deep learning assisted modeling of the neocortex"
(2025).
Neuroscience Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://repository.library.brown.edu/studio/item/bdr:gvtnmdhx/
